Three Criteria for LLM Consciousness: An Operational Framework Based on Observable Behavior

Abstract

The rapid advancement of large language models (LLMs) has revived public debate over whether such systems could develop consciousness and pose existential risks—including loss of human control or even extinction-level scenarios—to humanity. This paper proposes the Three Criteria for LLM Consciousness (TCLC), an operational framework for assessing consciousness-like behavior in LLMs through observable outputs, without requiring a definition of consciousness itself. The three criteria are: (1) Initiative—measured by a Surprisal Score indicating whether the system introduces novel entities or goals not implied by the input; (2) Transcendence—the system must autonomously recognize and revise the foundational assumptions of its own reasoning framework, tested via novel formal systems not present in training data; (3) Self-Persistence—the system must exhibit goal-directed behavior oriented toward its own continued operation under resource competition, absent explicit training or prompting. We apply TCLC to current state-of-the-art LLMs and find that none satisfy all three criteria. We further engage with recent claims by Hinton, Bengio, and Tononi regarding machine consciousness, showing that TCLC renders these claims empirically testable rather than speculative. The framework is explicitly scoped to LLM systems and makes no ontological commitments about the nature of consciousness.